A Cost-Aware Agentic Architecture for NL-to-SQL over Nested Enterprise Schemas, with a New Benchmark

On September 7, 2026, the researchers released the DevRev NL2SQL benchmark, which included 900 validated nested type and linked graph structure queries, along with a semantic depth scoring criterion. The team also proposed a cost-oriented single-round generative agent architecture, integrating pattern selection, metadata retrieval, and error repair components. In the DevRev NL2SQL benchmark, the system achieved a correct response rate of 91.7%, a 54.6 percentage point improvement over the suboptimal baseline; on the Spider 2.0 Snowflake public dataset, its single-round performance was comparable to that of leading systems.